From the first specialized Kashmiri LLM to a 28.5M-sample Indic OCR dataset and products with real users, I work where AI research meets software engineering.
CS Engineering student · AI Engineer · Full-Stack Builder
Four arXiv papers and an ongoing OCR research program, across NLP, computer vision, translation data and LLM evaluation.
One million sentences aligned across all 22 scheduled Indian languages, so any language pairs with any other: 462 directed pairs without direct translation data for each one. The translations are machine-generated; the paper proposes confidence and human-correction layers.
68 scenarios, 5 frontier LLMs, 1,800 API calls and a four-axis rubric. Every model handled single-sensor hazards but missed the ones that only appear when several sub-threshold readings combine.
613,078 synthetic image-text pairs for Kashmiri Nastaliq OCR, the first large-scale set for the script, built with SynthOCR-Gen and 25+ augmentation strategies across fonts, optical degradation and layout.
A ByT5-small model that restores diacritics in Kashmiri text (WER 0.2159, 77%+ native-speaker accuracy), with the first public dataset of 23.7K aligned diacritized sentence pairs. Covered by ETV Bharat.
Five OCR architecture generations over about 11 iterations on roughly 673K samples. The finding: the bottleneck is long-sequence decoding, not visual encoding. Active work is on v5.
Three datasets and six documented models. Some are gated, so you may need to accept the terms first.
Teach machines to read Indic scripts
5 releasesKashmiri to English and back
2 releasesPut the missing marks back
1 releaseA byte-level model trained from scratch
1 releaseNine systems, datasets and benchmarks. Each one starts from a real gap and ends in something you can open.
Kashmiri translation
Most Kashmiri writing is locked inside old InPage files. I unlocked it, taught a model the language, then tested it against Meta's translator. It scored higher in both directions.
Puts the marks back
Kashmiri is often typed without its small marks, so one spelling can mean several things. The model reads the raw bytes and writes the text back with the marks restored.
22 languages, any to any
One million sentences, each written in all 22 Indian languages and lined up row by row. So any language can be paired with any other: 462 directions from a single dataset.
28.5M OCR pairs
Real scans are messy. I render clean text across 71 languages, then wreck it in 54 different ways, so models learn to read the bad copies too.
5 model generations
I kept swapping in stronger models until it read better. The baseline copes with clean print and falls apart on old, dirty pages. That gap is the whole problem.
Synthetic pages
A page is built the way a print shop does it: pick the text, pick a font, lay it out, age it, check it. Over 600 thousand pages, none of them scanned by hand.
Do LLMs spot danger?
Each sensor alone looks safe. Add the three together and the room is in danger. I test whether AI models notice the danger that only shows up when readings add up.
Live photo rooms
Someone drops a photo into a room and everyone in it sees it instantly. No app, no sign-up. The server passes it along and the files come from a fast CDN.
Messy idea to clean prompts
You describe what you want in a messy sentence. It asks a few questions, then hands back clean, ordered prompts and sends each job to the model that suits it.
About 7 million people speak Kashmiri and almost no language technology existed for it. KoshurAI is a specialized Kashmiri-English translation model on Gemma 3: continued pretraining on a 2.8M-token Kashmiri corpus, then QLoRA fine-tuning on 16K+ bidirectional pairs. It scores higher COMET than Meta's NLLB-200-distilled-600M in both directions on FLORES-200. Presented at AI Impact Summit 2026.
Kashmiri is a low-resource language with near-zero representation in modern AI systems. There were no translation datasets, no NLP tools and no language model, which left a whole linguistic community outside the AI wave. Most source text also sits in InPage, a format that general tooling cannot read.
Kashmiri text often leaves out diacritics, which makes words ambiguous and hurts every downstream NLP task. Koshur Diacritizer is a ByT5-small seq2seq model that restores them. It is published on arXiv, released on Hugging Face and was covered by ETV Bharat.
The same undiacritized Kashmiri spelling can stand for several different words. Tokenizers, translation, search and text-to-speech all get worse without diacritics, and no neural solution existed for Kashmiri.
One million sentences, aligned line by line across all 22 scheduled Indian languages. Every row carries the same meaning in every language, so any language can be paired with any other: 462 directed pairs, with no need for direct translation data for each pair.
Most Indian languages have no direct translation data with each other, so something like Marathi to Kashmiri or Bodo to Dogri normally has to detour through English. BharatSetu builds the shared layer once, so every pair is available from the same table.
Co-authored with Haq Nawaz Malik and Nahfid Nissar. Translations are machine-generated, not human-verified; the paper proposes confidence and human-correction layers on top.
A synthetic document-OCR corpus covering 71 languages and 18 writing systems, from the 22 scheduled languages to regional, tribal and classical ones. Built and released by me on Hugging Face, with 3,500+ downloads.
Many Indian scripts have very little OCR training data, and Chakma, Lepcha, Limbu, Santali and Meetei Mayek have almost none. Real scans are costly to label and complex scripts need correct shaping, so BhartiOCR renders every sample through a proper text-shaping engine instead of pasting glyphs.
Synthetic data only. The card lists open limits: no separate validation or test shards yet, and lossy WebP images.
Five architecture generations over about 11 experimental iterations on roughly 673K training samples: the road from a CNN baseline to foundation-model OCR, and the discovery that redirected the whole program.
Each generation ruled out an architecture family: CNNs lacked sequence understanding, CTC plateaued on ligatures, PARSeq did not generalize to Nastaliq and TrOCR hit tokenizer limits. Custom evaluation (CER, WER, prediction coverage, decoder-length analysis) showed the real limit: decoder sequence collapse, not encoder capacity. The encoder learns Kashmiri glyphs; the decoder truncates long sequences.
The first large-scale synthetic OCR dataset for Kashmiri Nastaliq: 613,078 image-text pairs generated with the SynthOCR-Gen framework and published on arXiv. It is the data side of the KoshurOCR program.
OCR for low-resource scripts is limited by data more than by architecture. Hand-labelling Nastaliq does not scale, because contextual ligatures make character-level labels slow and error-prone. Synthetic generation is the only route to scale.
Can frontier LLMs read sensor data and reason correctly about physical danger? A 68-scenario benchmark of 5 frontier models across toxic gas, thermal, mechanical, intrusion and deliberately ambiguous cases, scored on four axes over 1,800 API calls.
LLMs are being wired into industrial monitoring, smart buildings and robot safety systems, yet no public benchmark tests whether they can turn multi-sensor readings into the right hazard call and the right action. Existing benchmarks test academic reasoning, not physical-world judgment.
Every model handled single-sensor hazards well but missed hazards that only appear when several sub-threshold readings add up.
A real-time, room-based photo sharing platform: no install, no sign-up, anonymous sessions and a gallery that updates live for everyone in the room. Supabase Realtime handles sync and Cloudflare R2 handles zero-egress media storage.
Group media tools are heavy, slow and built around accounts. Groopik is built for the moment at an event: scan, join a room, upload, see everyone's photos. Real users then exposed real problems, and each got a fix: stale-while-revalidate caching for CDN lag, client-side rate limiting for upload bursts, and chunked resumable uploads for weak event Wi-Fi.
A voice-first assistant that turns a vague request into a precise, structured prompt. It asks 3 to 5 clarifying questions first, then routes the job between Llama 3.3 70B and Claude Sonnet depending on the task.
Most people get weak answers from LLMs because the prompt is vague, not because the model is bad. PromptForge fixes quality at the input: voice or text in, clarification, then a decomposed set of ordered, self-contained prompts out.
Summits, national podiums, state titles and open-source work. Newest first.
From raw text to a live product: the tools, in the order I reach for them.
Turn text into millions of training images.
Teach models a language or a script.
Check them against real benchmarks.
Wrap the model in a product people use.
Store it, deploy it, keep it live.
Active engineering threads right now.
Fixing decoder sequence collapse in Nastaliq OCR with length-aware decoding objectives and short-to-long curriculum learning over the 673K-sample ecosystem.
ActiveRunning the first recognition benchmarks on the 28.5M-sample corpus (ResNet-BiLSTM-CTC on Hindi, Bengali, Tamil and Urdu) to separate the effect of degradation from typography and layout.
In progressSpeech and RAG over the Kashmiri corpus: the next piece of the Kashmiri AI stack, building on KoshurAI, the Diacritizer and OCR.
🔬 PrototypingContributing to AI and agent projects through GirlScript Summer of Code 2026, while improving Groopik and PromptForge from real user feedback.
ShippingI'm a builder from Kashmir who likes solving hard problems with AI, full-stack software and product engineering. I don't build things for the sake of building. Every project starts with a clear problem and ends with something measurable.
From the first specialized Kashmiri LLM to a 28.5M-sample Indic OCR dataset, four arXiv papers and products with real users, I focus on work that makes a measurable difference for real people.
I'm equally at home training a model on CUDA, designing a PCB for a combat robot, or architecting a backend. That range is what I bring to a team. I learn fast, adapt quickly, take the lead when it is needed and think a problem through before I touch the code.
The discipline that wins Wushu championships is the same discipline that ships code at 2am.
J&K State Level
National IIT Podiums
AI Summits & Events
15 AI Projects Shown
Archery · Football · Swim
Press and partnerships that came from the work.
Featured by Brut India as a university representative at AI Impact Summit 2026, presenting 15 AI projects.
"Three Kashmir Engineers Develop AI Tool to Preserve Native Language in Digital Age", covering our Kashmiri language AI work (KoshurAI and Koshur Diacritizer).
Coordinated robotics outreach with the Indian Army through the RISC club as Outreach Head, LPU.
30+ certificates from recognized platforms. A selection of the most relevant.
Let's talk.